Di Zhang, Xun Wu, Shaohan Huang, Lingjie Jiang Peking University, Yaru Hao Microsoft Research Asia, Li Dong Microsoft Research, Zewen Chi Microsoft Research, Zhifang Sui Peking University 等 9 位作者
Reinforcement learning with verifiable rewards (RLVR) has emerged as a powerful paradigm for improving reasoning capabilities. However, training RLVR with Mixture-of-Experts (MoE) policies remains fragile and is often prone to reward collapse.We identify a MoE-specific source of instability, referred to as router shift (RS), where changes in expert routing across policy updates exacerbate off-policy mismatch. This effect leads to increasingly volatile importance-ratio signals and bursty clipping behavior, which consistently precede training collapse.Motivated by this diagnosis, we propose Router-Shift Policy Optimization (RSPO). RSPO computes a per-token router-shift ratio conditioned on the previously activated experts, applies stop-gradient and a lower-bound floor, and softly rescales importance ratios prior to clipping and aggregation. This design explicitly accounts for routing-induced distributional drift during off-policy optimization.We evaluate the effect of RSPO under two settings: a synthetic countdown task and real-world reasoning tasks on MATH and Code. Across both settings, RSPO achieves better performance and exhibits greater stability compared to recent MoE-based RLVR methods.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Tianyi Wang, Yixia Li, Long Li Infly Tech, Yibiao Chen, Shaohan Huang, Yun Chen Shanghai University of Finance and Economics, Peng Li Tsinghua University, Yang Liu 等 9 位作者
Proximal Policy Optimization (PPO) is central to aligning Large Language Models (LLMs) in reasoning tasks with verifiable rewards. However, standard token-level PPO struggles in this setting due to the instability of temporal credit assignment over long Chain-of-Thought (CoT) horizons and the prohibitive memory cost of the value model. While critic-free alternatives like GRPO mitigate these issues, they incur significant computational overhead by requiring multiple samples for baseline estimation, severely limiting training throughput. In this paper, we introduce Sequence-Level PPO (SPPO), a scalable algorithm that harmonizes the sample efficiency of PPO with the stability of outcome-based updates. SPPO reformulates the reasoning process as a Sequence-Level Contextual Bandit problem, employing a decoupled scalar value function to derive low-variance advantage signals without multi-sampling. Extensive experiments on mathematical benchmarks demonstrate that SPPO significantly surpasses standard PPO and matches the performance of computation-heavy group-based methods, offering a resource-efficient framework for aligning reasoning LLMs.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Yixia Li, Yaqing Shi, Zhiwen Ruan, Dongdong Zhang Research, Microsoft, Lingjie Jiang Peking University, Shaohan Huang, Yun Chen Shanghai University of Finance and Economics, Guanhua Chen Southern University of Science and Technology 等 9 位作者
Multimodal large language models have advanced rapidly, yet most remain English-centric, as scaling multilingual multimodal instruction tuning is limited by the scarcity and high cost of high-quality non-English image–text supervision. Although multilingual text data is abundant, naive textual fine-tuning can disrupt vision–language alignment and induce catastrophic forgetting. We propose Vision-Free Adaptation (VFA), a framework that decouples multilingual language enhancement from visual alignment by composing complementary task vectors over a shared LLM backbone. Specifically, we fine-tune a base LLM on multilingual text data to derive a multilingual task vector, which is then merged with the vision-aligned task vector of an MLLM. Experiments on five MLLMs across six multilingual multimodal benchmarks show consistent improvements while preserving both general multimodal and text-only capabilities. Moreover, VFA attains competitive performance with a fully multimodally trained model using less than 2% of the text data, demonstrating its efficiency and effectiveness.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表